In the process of upgrading domestic processors toward high-performance computing, compilation optimization is key to fully unleashing the underlying computing power of chips and ensuring device adaptability and operational stability. For a long time, high-end compilation optimization technologies have had high technical thresholds and strong barriers, becoming a prominent weakness that restricts the performance of domestic chips and the improvement of their ecosystem.
Recently, Wu Xizhe, an undergraduate student at Dalian University of Technology, led the “Shuangqing Zhiyi” team in independently developing the OSCAR Intelligent Compilation Optimization System. The system has overcome core bottlenecks in compilation optimization for domestic chips, effectively improved the level of high-performance compilation for domestic processors, and provided a reliable technical solution for performance iteration and ecosystem construction of domestic chips. Based on the practical needs of industry, the team has continuously carried out intelligent optimization design, system development, and real-machine verification around the compilation pain points of domestic chip architectures in scenarios such as scientific computing, big data processing, and AI inference. “Only by thoroughly understanding the architectural characteristics of domestic processors can we build a truly practical compilation optimization system,” Wu Xizhe said.
In response to industry challenges in intelligent compilation optimization, including a broad optimization space, complex constraints, and difficult tuning, the team abandoned the traditional single heuristic optimization approach. Instead, it innovatively integrated adaptive optimization based on large models, intelligent swarm search algorithms, and a multi-level intelligent decision-making mechanism, building a multi-granularity system tuning framework for domestic processors. After hundreds of program tests and multiple rounds of model iteration, the team successfully completed system development and greatly improved the efficiency of exploring the underlying computing power of domestic chips.
With this achievement, Wu Xizhe, then in his sophomore year, participated in the 19th “Challenge Cup” Chinese Youth Science and Technology Innovation “Jiebang Guashuai” Competition. Verified on real machines, the system demonstrated significant performance improvements over traditional compilation baselines under typical computing workloads. It can intelligently explore optimization spaces and generate dedicated compilation schemes adapted to different application scenarios. The technical solution ultimately won the Special Prize for the competition topic, and the team was the only undergraduate team among the Special Prize winners.
During high school, team leader Wu Xizhe had already shown potential in the field of informatics competitions. He won the First Prize in the CSP-S Advanced Group and was selected for the Beijing Youth Science and Technology Reserve Talent Early Training Program. In addition, he participated in the “Turing Program” natural language processing course organized by Gaoling School of Artificial Intelligence at Renmin University of China, and won the Special Prize in the China Division of the 2021 International Mathematical Modeling Challenge (IMMC). After entering Dalian University of Technology in 2023, he joined Professor Jiang He’s research group at Dalian University of Technology, where he participated in the research and development of fuzz testing for an autonomous driving simulation testing system. He also received multiple honors, including Finalist, equivalent to a Special Prize nomination, in the Interdisciplinary Contest in Modeling (ICM) of the Mathematical Contest in Modeling (MCM). These competition experiences laid a foundation for him in algorithm design, code implementation, and team collaboration.
At present, the team has established Dalian Shuangqing Zhiyi Information Technology Co., Ltd. to promote technological iteration and industrial implementation. Going forward, the team will expand the adaptation of the OSCAR system to more domestic processor architectures, deepen integrated innovation between AI and compilation technologies, and provide precise services for domestic chip enterprises. From traditional manual tuning to intelligent autonomous tuning, this undergraduate team is using independent innovation to address weaknesses in domestic chip compilation technologies and inject youthful momentum into China’s efforts to overtake in high-performance domestic computing power.
Media Contact
Company Name: Shuangqingzhiyi
Contact Person: Wu Xizhe
Email: Send Email
City: Dalian
Country: China
Website: https://www.dlut.edu.cn
